kernel - Add usched_dfly algorith, set as default for now (8)
[dragonfly.git] / sys / kern / lwkt_thread.c
CommitLineData
8ad65e08 1/*
b12defdc 2 * Copyright (c) 2003-2011 The DragonFly Project. All rights reserved.
60f60350 3 *
8c10bfcf
MD
4 * This code is derived from software contributed to The DragonFly Project
5 * by Matthew Dillon <dillon@backplane.com>
60f60350 6 *
8ad65e08
MD
7 * Redistribution and use in source and binary forms, with or without
8 * modification, are permitted provided that the following conditions
9 * are met:
60f60350 10 *
8ad65e08
MD
11 * 1. Redistributions of source code must retain the above copyright
12 * notice, this list of conditions and the following disclaimer.
13 * 2. Redistributions in binary form must reproduce the above copyright
8c10bfcf
MD
14 * notice, this list of conditions and the following disclaimer in
15 * the documentation and/or other materials provided with the
16 * distribution.
17 * 3. Neither the name of The DragonFly Project nor the names of its
18 * contributors may be used to endorse or promote products derived
19 * from this software without specific, prior written permission.
60f60350 20 *
8c10bfcf
MD
21 * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
22 * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
23 * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS
24 * FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE
25 * COPYRIGHT HOLDERS OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT,
26 * INCIDENTAL, SPECIAL, EXEMPLARY OR CONSEQUENTIAL DAMAGES (INCLUDING,
27 * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
28 * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED
29 * AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
30 * OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT
31 * OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
8ad65e08 32 * SUCH DAMAGE.
75cdbe6c
MD
33 */
34
35/*
36 * Each cpu in a system has its own self-contained light weight kernel
37 * thread scheduler, which means that generally speaking we only need
38 * to use a critical section to avoid problems. Foreign thread
39 * scheduling is queued via (async) IPIs.
8ad65e08
MD
40 */
41
42#include <sys/param.h>
43#include <sys/systm.h>
44#include <sys/kernel.h>
45#include <sys/proc.h>
46#include <sys/rtprio.h>
b37f18d6 47#include <sys/kinfo.h>
8ad65e08 48#include <sys/queue.h>
7d0bac62 49#include <sys/sysctl.h>
99df837e 50#include <sys/kthread.h>
f1d1c3fa 51#include <machine/cpu.h>
99df837e 52#include <sys/lock.h>
f6bf3af1 53#include <sys/caps.h>
9d265729 54#include <sys/spinlock.h>
57aa743c 55#include <sys/ktr.h>
9d265729
MD
56
57#include <sys/thread2.h>
58#include <sys/spinlock2.h>
684a93c4 59#include <sys/mplock2.h>
f1d1c3fa 60
8c72e3d5
AH
61#include <sys/dsched.h>
62
7d0bac62
MD
63#include <vm/vm.h>
64#include <vm/vm_param.h>
65#include <vm/vm_kern.h>
66#include <vm/vm_object.h>
67#include <vm/vm_page.h>
68#include <vm/vm_map.h>
69#include <vm/vm_pager.h>
70#include <vm/vm_extern.h>
7d0bac62 71
99df837e 72#include <machine/stdarg.h>
96728c05 73#include <machine/smp.h>
99df837e 74
d850923c
AE
75#if !defined(KTR_CTXSW)
76#define KTR_CTXSW KTR_ALL
77#endif
78KTR_INFO_MASTER(ctxsw);
5bf48697
AE
79KTR_INFO(KTR_CTXSW, ctxsw, sw, 0, "#cpu[%d].td = %p", int cpu, struct thread *td);
80KTR_INFO(KTR_CTXSW, ctxsw, pre, 1, "#cpu[%d].td = %p", int cpu, struct thread *td);
81KTR_INFO(KTR_CTXSW, ctxsw, newtd, 2, "#threads[%p].name = %s", struct thread *td, char *comm);
82KTR_INFO(KTR_CTXSW, ctxsw, deadtd, 3, "#threads[%p].name = <dead>", struct thread *td);
1541028a 83
40aaf5fc
NT
84static MALLOC_DEFINE(M_THREAD, "thread", "lwkt threads");
85
0f7a3396
MD
86#ifdef INVARIANTS
87static int panic_on_cscount = 0;
88#endif
05220613
MD
89static __int64_t switch_count = 0;
90static __int64_t preempt_hit = 0;
91static __int64_t preempt_miss = 0;
92static __int64_t preempt_weird = 0;
85946b6c 93static __int64_t token_contention_count[TDPRI_MAX+1] __debugvar;
fb0f29c4 94static int lwkt_use_spin_port;
40aaf5fc 95static struct objcache *thread_cache;
05220613 96
88ebb169 97#ifdef SMP
e381e77c 98static void lwkt_schedule_remote(void *arg, int arg2, struct intrframe *frame);
cc9b6223 99static void lwkt_setcpu_remote(void *arg);
88ebb169 100#endif
e381e77c 101
0855a2af
JG
102extern void cpu_heavy_restore(void);
103extern void cpu_lwkt_restore(void);
104extern void cpu_kthread_restore(void);
105extern void cpu_idle_restore(void);
106
fb0f29c4
MD
107/*
108 * We can make all thread ports use the spin backend instead of the thread
109 * backend. This should only be set to debug the spin backend.
110 */
111TUNABLE_INT("lwkt.use_spin_port", &lwkt_use_spin_port);
112
0f7a3396 113#ifdef INVARIANTS
0c52fa62
SG
114SYSCTL_INT(_lwkt, OID_AUTO, panic_on_cscount, CTLFLAG_RW, &panic_on_cscount, 0,
115 "Panic if attempting to switch lwkt's while mastering cpusync");
0f7a3396 116#endif
0c52fa62
SG
117SYSCTL_QUAD(_lwkt, OID_AUTO, switch_count, CTLFLAG_RW, &switch_count, 0,
118 "Number of switched threads");
9733f757 119SYSCTL_QUAD(_lwkt, OID_AUTO, preempt_hit, CTLFLAG_RW, &preempt_hit, 0,
0c52fa62 120 "Successful preemption events");
9733f757 121SYSCTL_QUAD(_lwkt, OID_AUTO, preempt_miss, CTLFLAG_RW, &preempt_miss, 0,
0c52fa62
SG
122 "Failed preemption events");
123SYSCTL_QUAD(_lwkt, OID_AUTO, preempt_weird, CTLFLAG_RW, &preempt_weird, 0,
124 "Number of preempted threads.");
38717797 125#ifdef INVARIANTS
85946b6c
MD
126SYSCTL_QUAD(_lwkt, OID_AUTO, token_contention_count_00, CTLFLAG_RW,
127 &token_contention_count[0], 0, "spinning due to token contention");
128SYSCTL_QUAD(_lwkt, OID_AUTO, token_contention_count_01, CTLFLAG_RW,
129 &token_contention_count[1], 0, "spinning due to token contention");
130SYSCTL_QUAD(_lwkt, OID_AUTO, token_contention_count_02, CTLFLAG_RW,
131 &token_contention_count[2], 0, "spinning due to token contention");
132SYSCTL_QUAD(_lwkt, OID_AUTO, token_contention_count_03, CTLFLAG_RW,
133 &token_contention_count[3], 0, "spinning due to token contention");
134SYSCTL_QUAD(_lwkt, OID_AUTO, token_contention_count_04, CTLFLAG_RW,
135 &token_contention_count[4], 0, "spinning due to token contention");
136SYSCTL_QUAD(_lwkt, OID_AUTO, token_contention_count_05, CTLFLAG_RW,
137 &token_contention_count[5], 0, "spinning due to token contention");
138SYSCTL_QUAD(_lwkt, OID_AUTO, token_contention_count_06, CTLFLAG_RW,
139 &token_contention_count[6], 0, "spinning due to token contention");
140SYSCTL_QUAD(_lwkt, OID_AUTO, token_contention_count_07, CTLFLAG_RW,
141 &token_contention_count[7], 0, "spinning due to token contention");
142SYSCTL_QUAD(_lwkt, OID_AUTO, token_contention_count_08, CTLFLAG_RW,
143 &token_contention_count[8], 0, "spinning due to token contention");
144SYSCTL_QUAD(_lwkt, OID_AUTO, token_contention_count_09, CTLFLAG_RW,
145 &token_contention_count[9], 0, "spinning due to token contention");
146SYSCTL_QUAD(_lwkt, OID_AUTO, token_contention_count_10, CTLFLAG_RW,
147 &token_contention_count[10], 0, "spinning due to token contention");
148SYSCTL_QUAD(_lwkt, OID_AUTO, token_contention_count_11, CTLFLAG_RW,
149 &token_contention_count[11], 0, "spinning due to token contention");
150SYSCTL_QUAD(_lwkt, OID_AUTO, token_contention_count_12, CTLFLAG_RW,
151 &token_contention_count[12], 0, "spinning due to token contention");
152SYSCTL_QUAD(_lwkt, OID_AUTO, token_contention_count_13, CTLFLAG_RW,
153 &token_contention_count[13], 0, "spinning due to token contention");
154SYSCTL_QUAD(_lwkt, OID_AUTO, token_contention_count_14, CTLFLAG_RW,
155 &token_contention_count[14], 0, "spinning due to token contention");
156SYSCTL_QUAD(_lwkt, OID_AUTO, token_contention_count_15, CTLFLAG_RW,
157 &token_contention_count[15], 0, "spinning due to token contention");
158SYSCTL_QUAD(_lwkt, OID_AUTO, token_contention_count_16, CTLFLAG_RW,
159 &token_contention_count[16], 0, "spinning due to token contention");
160SYSCTL_QUAD(_lwkt, OID_AUTO, token_contention_count_17, CTLFLAG_RW,
161 &token_contention_count[17], 0, "spinning due to token contention");
162SYSCTL_QUAD(_lwkt, OID_AUTO, token_contention_count_18, CTLFLAG_RW,
163 &token_contention_count[18], 0, "spinning due to token contention");
164SYSCTL_QUAD(_lwkt, OID_AUTO, token_contention_count_19, CTLFLAG_RW,
165 &token_contention_count[19], 0, "spinning due to token contention");
166SYSCTL_QUAD(_lwkt, OID_AUTO, token_contention_count_20, CTLFLAG_RW,
167 &token_contention_count[20], 0, "spinning due to token contention");
168SYSCTL_QUAD(_lwkt, OID_AUTO, token_contention_count_21, CTLFLAG_RW,
169 &token_contention_count[21], 0, "spinning due to token contention");
170SYSCTL_QUAD(_lwkt, OID_AUTO, token_contention_count_22, CTLFLAG_RW,
171 &token_contention_count[22], 0, "spinning due to token contention");
172SYSCTL_QUAD(_lwkt, OID_AUTO, token_contention_count_23, CTLFLAG_RW,
173 &token_contention_count[23], 0, "spinning due to token contention");
174SYSCTL_QUAD(_lwkt, OID_AUTO, token_contention_count_24, CTLFLAG_RW,
175 &token_contention_count[24], 0, "spinning due to token contention");
176SYSCTL_QUAD(_lwkt, OID_AUTO, token_contention_count_25, CTLFLAG_RW,
177 &token_contention_count[25], 0, "spinning due to token contention");
178SYSCTL_QUAD(_lwkt, OID_AUTO, token_contention_count_26, CTLFLAG_RW,
179 &token_contention_count[26], 0, "spinning due to token contention");
180SYSCTL_QUAD(_lwkt, OID_AUTO, token_contention_count_27, CTLFLAG_RW,
181 &token_contention_count[27], 0, "spinning due to token contention");
182SYSCTL_QUAD(_lwkt, OID_AUTO, token_contention_count_28, CTLFLAG_RW,
183 &token_contention_count[28], 0, "spinning due to token contention");
184SYSCTL_QUAD(_lwkt, OID_AUTO, token_contention_count_29, CTLFLAG_RW,
185 &token_contention_count[29], 0, "spinning due to token contention");
186SYSCTL_QUAD(_lwkt, OID_AUTO, token_contention_count_30, CTLFLAG_RW,
187 &token_contention_count[30], 0, "spinning due to token contention");
188SYSCTL_QUAD(_lwkt, OID_AUTO, token_contention_count_31, CTLFLAG_RW,
189 &token_contention_count[31], 0, "spinning due to token contention");
38717797 190#endif
b12defdc 191static int fairq_enable = 0;
2a418930
MD
192SYSCTL_INT(_lwkt, OID_AUTO, fairq_enable, CTLFLAG_RW,
193 &fairq_enable, 0, "Turn on fairq priority accumulators");
85946b6c 194static int fairq_bypass = -1;
b12defdc
MD
195SYSCTL_INT(_lwkt, OID_AUTO, fairq_bypass, CTLFLAG_RW,
196 &fairq_bypass, 0, "Allow fairq to bypass td on token failure");
197extern int lwkt_sched_debug;
198int lwkt_sched_debug = 0;
199SYSCTL_INT(_lwkt, OID_AUTO, sched_debug, CTLFLAG_RW,
200 &lwkt_sched_debug, 0, "Scheduler debug");
2a418930
MD
201static int lwkt_spin_loops = 10;
202SYSCTL_INT(_lwkt, OID_AUTO, spin_loops, CTLFLAG_RW,
b12defdc
MD
203 &lwkt_spin_loops, 0, "Scheduler spin loops until sorted decon");
204static int lwkt_spin_reseq = 0;
205SYSCTL_INT(_lwkt, OID_AUTO, spin_reseq, CTLFLAG_RW,
206 &lwkt_spin_reseq, 0, "Scheduler resequencer enable");
207static int lwkt_spin_monitor = 0;
208SYSCTL_INT(_lwkt, OID_AUTO, spin_monitor, CTLFLAG_RW,
209 &lwkt_spin_monitor, 0, "Scheduler uses monitor/mwait");
d5b2d319
MD
210static int lwkt_spin_fatal = 0; /* disabled */
211SYSCTL_INT(_lwkt, OID_AUTO, spin_fatal, CTLFLAG_RW,
212 &lwkt_spin_fatal, 0, "LWKT scheduler spin loops till fatal panic");
fbc024e4 213static int preempt_enable = 1;
2a418930
MD
214SYSCTL_INT(_lwkt, OID_AUTO, preempt_enable, CTLFLAG_RW,
215 &preempt_enable, 0, "Enable preemption");
7b234d8c 216static int lwkt_cache_threads = 0;
765b1ae0
MD
217SYSCTL_INT(_lwkt, OID_AUTO, cache_threads, CTLFLAG_RD,
218 &lwkt_cache_threads, 0, "thread+kstack cache");
fbc024e4 219
2a418930
MD
220static __cachealign int lwkt_cseq_rindex;
221static __cachealign int lwkt_cseq_windex;
05220613 222
4b5f931b
MD
223/*
224 * These helper procedures handle the runq, they can only be called from
225 * within a critical section.
75cdbe6c
MD
226 *
227 * WARNING! Prior to SMP being brought up it is possible to enqueue and
228 * dequeue threads belonging to other cpus, so be sure to use td->td_gd
229 * instead of 'mycpu' when referencing the globaldata structure. Once
230 * SMP live enqueuing and dequeueing only occurs on the current cpu.
4b5f931b 231 */
f1d1c3fa
MD
232static __inline
233void
234_lwkt_dequeue(thread_t td)
235{
236 if (td->td_flags & TDF_RUNQ) {
75cdbe6c 237 struct globaldata *gd = td->td_gd;
4b5f931b 238
f1d1c3fa 239 td->td_flags &= ~TDF_RUNQ;
f9235b6d 240 TAILQ_REMOVE(&gd->gd_tdrunq, td, td_threadq);
f9235b6d 241 if (TAILQ_FIRST(&gd->gd_tdrunq) == NULL)
2a418930 242 atomic_clear_int(&gd->gd_reqflags, RQF_RUNNING);
f1d1c3fa
MD
243 }
244}
245
f9235b6d
MD
246/*
247 * Priority enqueue.
248 *
249 * NOTE: There are a limited number of lwkt threads runnable since user
250 * processes only schedule one at a time per cpu.
251 */
f1d1c3fa
MD
252static __inline
253void
254_lwkt_enqueue(thread_t td)
255{
f9235b6d
MD
256 thread_t xtd;
257
7f5d7ed7 258 if ((td->td_flags & (TDF_RUNQ|TDF_MIGRATING|TDF_BLOCKQ)) == 0) {
75cdbe6c 259 struct globaldata *gd = td->td_gd;
4b5f931b 260
f1d1c3fa 261 td->td_flags |= TDF_RUNQ;
f9235b6d
MD
262 xtd = TAILQ_FIRST(&gd->gd_tdrunq);
263 if (xtd == NULL) {
85946b6c
MD
264 TAILQ_INSERT_TAIL(&gd->gd_tdrunq, td, td_threadq);
265 atomic_set_int(&gd->gd_reqflags, RQF_RUNNING);
f9235b6d 266 } else {
85946b6c
MD
267 while (xtd && xtd->td_pri >= td->td_pri)
268 xtd = TAILQ_NEXT(xtd, td_threadq);
269 if (xtd)
270 TAILQ_INSERT_BEFORE(xtd, td, td_threadq);
271 else
272 TAILQ_INSERT_TAIL(&gd->gd_tdrunq, td, td_threadq);
f9235b6d 273 }
b12defdc
MD
274
275 /*
85946b6c 276 * Request a LWKT reschedule if we are now at the head of the queue.
b12defdc 277 */
85946b6c
MD
278 if (TAILQ_FIRST(&gd->gd_tdrunq) == td)
279 need_lwkt_resched();
f1d1c3fa
MD
280 }
281}
8ad65e08 282
40aaf5fc
NT
283static __boolean_t
284_lwkt_thread_ctor(void *obj, void *privdata, int ocflags)
285{
286 struct thread *td = (struct thread *)obj;
287
288 td->td_kstack = NULL;
289 td->td_kstack_size = 0;
290 td->td_flags = TDF_ALLOCATED_THREAD;
4643740a 291 td->td_mpflags = 0;
40aaf5fc
NT
292 return (1);
293}
294
295static void
296_lwkt_thread_dtor(void *obj, void *privdata)
297{
298 struct thread *td = (struct thread *)obj;
299
300 KASSERT(td->td_flags & TDF_ALLOCATED_THREAD,
301 ("_lwkt_thread_dtor: not allocated from objcache"));
302 KASSERT((td->td_flags & TDF_ALLOCATED_STACK) && td->td_kstack &&
303 td->td_kstack_size > 0,
304 ("_lwkt_thread_dtor: corrupted stack"));
305 kmem_free(&kernel_map, (vm_offset_t)td->td_kstack, td->td_kstack_size);
7b234d8c
MD
306 td->td_kstack = NULL;
307 td->td_flags = 0;
40aaf5fc
NT
308}
309
310/*
311 * Initialize the lwkt s/system.
765b1ae0 312 *
7b234d8c
MD
313 * Nominally cache up to 32 thread + kstack structures. Cache more on
314 * systems with a lot of cpu cores.
40aaf5fc
NT
315 */
316void
317lwkt_init(void)
318{
765b1ae0 319 TUNABLE_INT("lwkt.cache_threads", &lwkt_cache_threads);
7b234d8c
MD
320 if (lwkt_cache_threads == 0) {
321 lwkt_cache_threads = ncpus * 4;
322 if (lwkt_cache_threads < 32)
323 lwkt_cache_threads = 32;
324 }
765b1ae0
MD
325 thread_cache = objcache_create_mbacked(
326 M_THREAD, sizeof(struct thread),
327 NULL, lwkt_cache_threads,
328 _lwkt_thread_ctor, _lwkt_thread_dtor, NULL);
40aaf5fc
NT
329}
330
37af14fe
MD
331/*
332 * Schedule a thread to run. As the current thread we can always safely
333 * schedule ourselves, and a shortcut procedure is provided for that
334 * function.
335 *
336 * (non-blocking, self contained on a per cpu basis)
337 */
338void
339lwkt_schedule_self(thread_t td)
340{
cfaeae2a 341 KKASSERT((td->td_flags & TDF_MIGRATING) == 0);
37af14fe 342 crit_enter_quick(td);
f9235b6d
MD
343 KASSERT(td != &td->td_gd->gd_idlethread,
344 ("lwkt_schedule_self(): scheduling gd_idlethread is illegal!"));
4643740a
MD
345 KKASSERT(td->td_lwp == NULL ||
346 (td->td_lwp->lwp_mpflags & LWP_MP_ONRUNQ) == 0);
37af14fe 347 _lwkt_enqueue(td);
37af14fe
MD
348 crit_exit_quick(td);
349}
350
351/*
352 * Deschedule a thread.
353 *
354 * (non-blocking, self contained on a per cpu basis)
355 */
356void
357lwkt_deschedule_self(thread_t td)
358{
359 crit_enter_quick(td);
37af14fe
MD
360 _lwkt_dequeue(td);
361 crit_exit_quick(td);
362}
363
8ad65e08
MD
364/*
365 * LWKTs operate on a per-cpu basis
366 *
73e4f7b9 367 * WARNING! Called from early boot, 'mycpu' may not work yet.
8ad65e08
MD
368 */
369void
370lwkt_gdinit(struct globaldata *gd)
371{
f9235b6d 372 TAILQ_INIT(&gd->gd_tdrunq);
73e4f7b9 373 TAILQ_INIT(&gd->gd_tdallq);
8ad65e08
MD
374}
375
7d0bac62
MD
376/*
377 * Create a new thread. The thread must be associated with a process context
75cdbe6c
MD
378 * or LWKT start address before it can be scheduled. If the target cpu is
379 * -1 the thread will be created on the current cpu.
0cfcada1
MD
380 *
381 * If you intend to create a thread without a process context this function
382 * does everything except load the startup and switcher function.
7d0bac62
MD
383 */
384thread_t
d3d32139 385lwkt_alloc_thread(struct thread *td, int stksize, int cpu, int flags)
7d0bac62 386{
d2d8515b 387 static int cpu_rotator;
c070746a 388 globaldata_t gd = mycpu;
99df837e 389 void *stack;
7d0bac62 390
c070746a
MD
391 /*
392 * If static thread storage is not supplied allocate a thread. Reuse
393 * a cached free thread if possible. gd_freetd is used to keep an exiting
394 * thread intact through the exit.
395 */
ef0fdad1 396 if (td == NULL) {
cf709dd2
MD
397 crit_enter_gd(gd);
398 if ((td = gd->gd_freetd) != NULL) {
399 KKASSERT((td->td_flags & (TDF_RUNNING|TDF_PREEMPT_LOCK|
400 TDF_RUNQ)) == 0);
c070746a 401 gd->gd_freetd = NULL;
cf709dd2 402 } else {
c070746a 403 td = objcache_get(thread_cache, M_WAITOK);
cf709dd2
MD
404 KKASSERT((td->td_flags & (TDF_RUNNING|TDF_PREEMPT_LOCK|
405 TDF_RUNQ)) == 0);
406 }
407 crit_exit_gd(gd);
40aaf5fc 408 KASSERT((td->td_flags &
2af9d75d
MD
409 (TDF_ALLOCATED_THREAD|TDF_RUNNING|TDF_PREEMPT_LOCK)) ==
410 TDF_ALLOCATED_THREAD,
40aaf5fc
NT
411 ("lwkt_alloc_thread: corrupted td flags 0x%X", td->td_flags));
412 flags |= td->td_flags & (TDF_ALLOCATED_THREAD|TDF_ALLOCATED_STACK);
ef0fdad1 413 }
c070746a
MD
414
415 /*
416 * Try to reuse cached stack.
417 */
f470d0c8
MD
418 if ((stack = td->td_kstack) != NULL && td->td_kstack_size != stksize) {
419 if (flags & TDF_ALLOCATED_STACK) {
e4846942 420 kmem_free(&kernel_map, (vm_offset_t)stack, td->td_kstack_size);
f470d0c8
MD
421 stack = NULL;
422 }
423 }
424 if (stack == NULL) {
e40cfbd7 425 stack = (void *)kmem_alloc_stack(&kernel_map, stksize);
ef0fdad1 426 flags |= TDF_ALLOCATED_STACK;
99df837e 427 }
d2d8515b
MD
428 if (cpu < 0) {
429 cpu = ++cpu_rotator;
430 cpu_ccfence();
431 cpu %= ncpus;
432 }
433 lwkt_init_thread(td, stack, stksize, flags, globaldata_find(cpu));
99df837e 434 return(td);
7d0bac62
MD
435}
436
437/*
438 * Initialize a preexisting thread structure. This function is used by
439 * lwkt_alloc_thread() and also used to initialize the per-cpu idlethread.
440 *
f8c3996b
MD
441 * All threads start out in a critical section at a priority of
442 * TDPRI_KERN_DAEMON. Higher level code will modify the priority as
75cdbe6c
MD
443 * appropriate. This function may send an IPI message when the
444 * requested cpu is not the current cpu and consequently gd_tdallq may
445 * not be initialized synchronously from the point of view of the originating
446 * cpu.
447 *
448 * NOTE! we have to be careful in regards to creating threads for other cpus
449 * if SMP has not yet been activated.
7d0bac62 450 */
41a01a4d
MD
451#ifdef SMP
452
75cdbe6c
MD
453static void
454lwkt_init_thread_remote(void *arg)
455{
456 thread_t td = arg;
457
52eedfb5
MD
458 /*
459 * Protected by critical section held by IPI dispatch
460 */
75cdbe6c
MD
461 TAILQ_INSERT_TAIL(&td->td_gd->gd_tdallq, td, td_allq);
462}
463
41a01a4d
MD
464#endif
465
fdce8919
MD
466/*
467 * lwkt core thread structural initialization.
468 *
469 * NOTE: All threads are initialized as mpsafe threads.
470 */
7d0bac62 471void
f470d0c8
MD
472lwkt_init_thread(thread_t td, void *stack, int stksize, int flags,
473 struct globaldata *gd)
7d0bac62 474{
37af14fe
MD
475 globaldata_t mygd = mycpu;
476
99df837e
MD
477 bzero(td, sizeof(struct thread));
478 td->td_kstack = stack;
f470d0c8 479 td->td_kstack_size = stksize;
d3d32139 480 td->td_flags = flags;
4643740a 481 td->td_mpflags = 0;
26a0694b 482 td->td_gd = gd;
f9235b6d
MD
483 td->td_pri = TDPRI_KERN_DAEMON;
484 td->td_critcount = 1;
54341a3b 485 td->td_toks_have = NULL;
3b998fa9 486 td->td_toks_stop = &td->td_toks_base;
392cd266 487 if (lwkt_use_spin_port || (flags & TDF_FORCE_SPINPORT))
d636cd48 488 lwkt_initport_spin(&td->td_msgport, td);
fb0f29c4
MD
489 else
490 lwkt_initport_thread(&td->td_msgport, td);
99df837e 491 pmap_init_thread(td);
0f7a3396 492#ifdef SMP
5d21b981
MD
493 /*
494 * Normally initializing a thread for a remote cpu requires sending an
495 * IPI. However, the idlethread is setup before the other cpus are
496 * activated so we have to treat it as a special case. XXX manipulation
497 * of gd_tdallq requires the BGL.
498 */
499 if (gd == mygd || td == &gd->gd_idlethread) {
37af14fe 500 crit_enter_gd(mygd);
75cdbe6c 501 TAILQ_INSERT_TAIL(&gd->gd_tdallq, td, td_allq);
37af14fe 502 crit_exit_gd(mygd);
75cdbe6c 503 } else {
2db3b277 504 lwkt_send_ipiq(gd, lwkt_init_thread_remote, td);
75cdbe6c 505 }
0f7a3396 506#else
37af14fe 507 crit_enter_gd(mygd);
0f7a3396 508 TAILQ_INSERT_TAIL(&gd->gd_tdallq, td, td_allq);
37af14fe 509 crit_exit_gd(mygd);
0f7a3396 510#endif
8c72e3d5
AH
511
512 dsched_new_thread(td);
73e4f7b9
MD
513}
514
515void
516lwkt_set_comm(thread_t td, const char *ctl, ...)
517{
e2565a42 518 __va_list va;
73e4f7b9 519
e2565a42 520 __va_start(va, ctl);
379210cb 521 kvsnprintf(td->td_comm, sizeof(td->td_comm), ctl, va);
e2565a42 522 __va_end(va);
5bf48697 523 KTR_LOG(ctxsw_newtd, td, td->td_comm);
7d0bac62
MD
524}
525
eb2adbf5
MD
526/*
527 * Prevent the thread from getting destroyed. Note that unlike PHOLD/PRELE
528 * this does not prevent the thread from migrating to another cpu so the
529 * gd_tdallq state is not protected by this.
530 */
99df837e 531void
73e4f7b9 532lwkt_hold(thread_t td)
99df837e 533{
74c9628e 534 atomic_add_int(&td->td_refs, 1);
73e4f7b9
MD
535}
536
537void
538lwkt_rele(thread_t td)
539{
540 KKASSERT(td->td_refs > 0);
74c9628e 541 atomic_add_int(&td->td_refs, -1);
73e4f7b9
MD
542}
543
73e4f7b9
MD
544void
545lwkt_free_thread(thread_t td)
546{
74c9628e 547 KKASSERT(td->td_refs == 0);
c17a6852
MD
548 KKASSERT((td->td_flags & (TDF_RUNNING | TDF_PREEMPT_LOCK |
549 TDF_RUNQ | TDF_TSLEEPQ)) == 0);
40aaf5fc
NT
550 if (td->td_flags & TDF_ALLOCATED_THREAD) {
551 objcache_put(thread_cache, td);
552 } else if (td->td_flags & TDF_ALLOCATED_STACK) {
553 /* client-allocated struct with internally allocated stack */
554 KASSERT(td->td_kstack && td->td_kstack_size > 0,
555 ("lwkt_free_thread: corrupted stack"));
556 kmem_free(&kernel_map, (vm_offset_t)td->td_kstack, td->td_kstack_size);
557 td->td_kstack = NULL;
558 td->td_kstack_size = 0;
99df837e 559 }
e7c0dbba 560 KTR_LOG(ctxsw_deadtd, td);
99df837e
MD
561}
562
563
8ad65e08
MD
564/*
565 * Switch to the next runnable lwkt. If no LWKTs are runnable then
f1d1c3fa
MD
566 * switch to the idlethread. Switching must occur within a critical
567 * section to avoid races with the scheduling queue.
568 *
569 * We always have full control over our cpu's run queue. Other cpus
570 * that wish to manipulate our queue must use the cpu_*msg() calls to
571 * talk to our cpu, so a critical section is all that is needed and
572 * the result is very, very fast thread switching.
573 *
96728c05
MD
574 * The LWKT scheduler uses a fixed priority model and round-robins at
575 * each priority level. User process scheduling is a totally
576 * different beast and LWKT priorities should not be confused with
577 * user process priorities.
f1d1c3fa 578 *
69d78e99
MD
579 * PREEMPTION NOTE: Preemption occurs via lwkt_preempt(). lwkt_switch()
580 * is not called by the current thread in the preemption case, only when
581 * the preempting thread blocks (in order to return to the original thread).
cfaeae2a
MD
582 *
583 * SPECIAL NOTE ON SWITCH ATOMICY: Certain operations such as thread
584 * migration and tsleep deschedule the current lwkt thread and call
585 * lwkt_switch(). In particular, the target cpu of the migration fully
586 * expects the thread to become non-runnable and can deadlock against
587 * cpusync operations if we run any IPIs prior to switching the thread out.
588 *
589 * WE MUST BE VERY CAREFUL NOT TO RUN SPLZ DIRECTLY OR INDIRECTLY IF
95858b91 590 * THE CURRENT THREAD HAS BEEN DESCHEDULED!
8ad65e08
MD
591 */
592void
593lwkt_switch(void)
594{
37af14fe
MD
595 globaldata_t gd = mycpu;
596 thread_t td = gd->gd_curthread;
8ad65e08 597 thread_t ntd;
b12defdc 598 int spinning = 0;
8ad65e08 599
da0b0e8b 600 KKASSERT(gd->gd_processing_ipiq == 0);
121f93bc 601 KKASSERT(td->td_flags & TDF_RUNNING);
da0b0e8b 602
46a3f46d 603 /*
27e88a6e
MD
604 * Switching from within a 'fast' (non thread switched) interrupt or IPI
605 * is illegal. However, we may have to do it anyway if we hit a fatal
606 * kernel trap or we have paniced.
607 *
608 * If this case occurs save and restore the interrupt nesting level.
46a3f46d 609 */
27e88a6e
MD
610 if (gd->gd_intr_nesting_level) {
611 int savegdnest;
612 int savegdtrap;
613
5fddbda2 614 if (gd->gd_trap_nesting_level == 0 && panic_cpu_gd != mycpu) {
4a28fe22 615 panic("lwkt_switch: Attempt to switch from a "
5a8df152 616 "fast interrupt, ipi, or hard code section, "
4a28fe22
MD
617 "td %p\n",
618 td);
27e88a6e
MD
619 } else {
620 savegdnest = gd->gd_intr_nesting_level;
621 savegdtrap = gd->gd_trap_nesting_level;
622 gd->gd_intr_nesting_level = 0;
623 gd->gd_trap_nesting_level = 0;
a7422615
MD
624 if ((td->td_flags & TDF_PANICWARN) == 0) {
625 td->td_flags |= TDF_PANICWARN;
4a28fe22
MD
626 kprintf("Warning: thread switch from interrupt, IPI, "
627 "or hard code section.\n"
a7422615 628 "thread %p (%s)\n", td, td->td_comm);
7ce2998e 629 print_backtrace(-1);
a7422615 630 }
27e88a6e
MD
631 lwkt_switch();
632 gd->gd_intr_nesting_level = savegdnest;
633 gd->gd_trap_nesting_level = savegdtrap;
634 return;
635 }
96728c05 636 }
ef0fdad1 637
cb973d15 638 /*
85946b6c
MD
639 * Release our current user process designation if we are blocking
640 * or if a user reschedule was requested.
641 *
642 * NOTE: This function is NOT called if we are switching into or
643 * returning from a preemption.
644 *
645 * NOTE: Releasing our current user process designation may cause
646 * it to be assigned to another thread, which in turn will
647 * cause us to block in the usched acquire code when we attempt
648 * to return to userland.
649 *
650 * NOTE: On SMP systems this can be very nasty when heavy token
651 * contention is present so we want to be careful not to
652 * release the designation gratuitously.
cb973d15 653 */
85946b6c
MD
654 if (td->td_release &&
655 (user_resched_wanted() || (td->td_flags & TDF_RUNQ) == 0)) {
cb973d15 656 td->td_release(td);
85946b6c 657 }
cb973d15 658
85946b6c
MD
659 /*
660 * Release all tokens
661 */
37af14fe 662 crit_enter_gd(gd);
3b998fa9 663 if (TD_TOKS_HELD(td))
9d265729
MD
664 lwkt_relalltokens(td);
665
666 /*
b02926de
MD
667 * We had better not be holding any spin locks, but don't get into an
668 * endless panic loop.
9d265729 669 */
d666840a
MD
670 KASSERT(gd->gd_spinlocks_wr == 0 || panicstr != NULL,
671 ("lwkt_switch: still holding %d exclusive spinlocks!",
672 gd->gd_spinlocks_wr));
9d265729 673
8a8d5d85
MD
674
675#ifdef SMP
0f7a3396
MD
676#ifdef INVARIANTS
677 if (td->td_cscount) {
6ea70f76 678 kprintf("Diagnostic: attempt to switch while mastering cpusync: %p\n",
0f7a3396
MD
679 td);
680 if (panic_on_cscount)
681 panic("switching while mastering cpusync");
682 }
683#endif
8a8d5d85 684#endif
f9235b6d
MD
685
686 /*
687 * If we had preempted another thread on this cpu, resume the preempted
688 * thread. This occurs transparently, whether the preempted thread
689 * was scheduled or not (it may have been preempted after descheduling
690 * itself).
691 *
692 * We have to setup the MP lock for the original thread after backing
693 * out the adjustment that was made to curthread when the original
694 * was preempted.
695 */
99df837e 696 if ((ntd = td->td_preempted) != NULL) {
26a0694b
MD
697 KKASSERT(ntd->td_flags & TDF_PREEMPT_LOCK);
698 ntd->td_flags |= TDF_PREEMPT_DONE;
8ec60c3f
MD
699
700 /*
b9eb1c19
MD
701 * The interrupt may have woken a thread up, we need to properly
702 * set the reschedule flag if the originally interrupted thread is
703 * at a lower priority.
85946b6c
MD
704 *
705 * The interrupt may not have descheduled.
8ec60c3f 706 */
85946b6c 707 if (TAILQ_FIRST(&gd->gd_tdrunq) != ntd)
8ec60c3f 708 need_lwkt_resched();
f9235b6d
MD
709 goto havethread_preempted;
710 }
711
b12defdc 712 /*
f9235b6d 713 * If we cannot obtain ownership of the tokens we cannot immediately
cfaeae2a
MD
714 * schedule the target thread.
715 *
716 * Reminder: Again, we cannot afford to run any IPIs in this path if
717 * the current thread has been descheduled.
f9235b6d
MD
718 */
719 for (;;) {
b12defdc 720 clear_lwkt_resched();
f9235b6d 721
4b5f931b 722 /*
2a418930 723 * Hotpath - pull the head of the run queue and attempt to schedule
85946b6c 724 * it.
41a01a4d 725 */
7c3dd309
NA
726 ntd = TAILQ_FIRST(&gd->gd_tdrunq);
727
728 if (ntd == NULL) {
729 /*
730 * Runq is empty, switch to idle to allow it to halt.
731 */
732 ntd = &gd->gd_idlethread;
6f207a2c 733#ifdef SMP
7c3dd309
NA
734 if (gd->gd_trap_nesting_level == 0 && panicstr == NULL)
735 ASSERT_NO_TOKENS_HELD(ntd);
6f207a2c 736#endif
7c3dd309
NA
737 cpu_time.cp_msg[0] = 0;
738 cpu_time.cp_stallpc = 0;
739 goto haveidle;
f9235b6d 740 }
41a01a4d 741
8ec60c3f 742 /*
b12defdc 743 * Hotpath - schedule ntd.
6f207a2c
MD
744 *
745 * NOTE: For UP there is no mplock and lwkt_getalltokens()
746 * always succeeds.
8ec60c3f 747 */
b12defdc
MD
748 if (TD_TOKS_NOT_HELD(ntd) ||
749 lwkt_getalltokens(ntd, (spinning >= lwkt_spin_loops)))
750 {
f9235b6d 751 goto havethread;
b12defdc 752 }
f9235b6d 753
f9235b6d 754 /*
2a418930
MD
755 * Coldpath (SMP only since tokens always succeed on UP)
756 *
757 * We had some contention on the thread we wanted to schedule.
758 * What we do now is try to find a thread that we can schedule
b12defdc 759 * in its stead.
2a418930 760 *
85946b6c
MD
761 * The coldpath scan does NOT rearrange threads in the run list.
762 * The lwkt_schedulerclock() will assert need_lwkt_resched() on
763 * the next tick whenever the current head is not the current thread.
f9235b6d 764 */
b12defdc 765#ifdef INVARIANTS
85946b6c
MD
766 ++token_contention_count[ntd->td_pri];
767 ++ntd->td_contended;
b12defdc
MD
768#endif
769
85946b6c 770 if (fairq_bypass > 0)
b12defdc
MD
771 goto skip;
772
b12defdc 773 while ((ntd = TAILQ_NEXT(ntd, td_threadq)) != NULL) {
85946b6c
MD
774 /*
775 * Never schedule threads returning to userland or the
776 * user thread scheduler helper thread when higher priority
bb6860b5
NA
777 * threads are present. The runq is sorted by priority
778 * so we can give up traversing it when we find the first
779 * low priority thread.
85946b6c
MD
780 */
781 if (ntd->td_pri < TDPRI_KERN_LPSCHED) {
782 ntd = NULL;
783 break;
784 }
785
786 /*
787 * Try this one.
788 */
b12defdc
MD
789 if (TD_TOKS_NOT_HELD(ntd) ||
790 lwkt_getalltokens(ntd, (spinning >= lwkt_spin_loops))) {
791 goto havethread;
792 }
85946b6c
MD
793#ifdef INVARIANTS
794 ++token_contention_count[ntd->td_pri];
795 ++ntd->td_contended;
b12defdc 796#endif
2a418930
MD
797 }
798
b12defdc 799skip:
2a418930
MD
800 /*
801 * We exhausted the run list, meaning that all runnable threads
b12defdc 802 * are contested.
2a418930
MD
803 */
804 cpu_pause();
805 ntd = &gd->gd_idlethread;
806#ifdef SMP
807 if (gd->gd_trap_nesting_level == 0 && panicstr == NULL)
808 ASSERT_NO_TOKENS_HELD(ntd);
809 /* contention case, do not clear contention mask */
810#endif
811
812 /*
b12defdc
MD
813 * We are going to have to retry but if the current thread is not
814 * on the runq we instead switch through the idle thread to get away
815 * from the current thread. We have to flag for lwkt reschedule
816 * to prevent the idle thread from halting.
2a418930 817 *
b12defdc
MD
818 * NOTE: A non-zero spinning is passed to lwkt_getalltokens() to
819 * instruct it to deal with the potential for deadlocks by
820 * ordering the tokens by address.
2a418930 821 */
b12defdc 822 if ((td->td_flags & TDF_RUNQ) == 0) {
85946b6c 823 need_lwkt_resched(); /* prevent hlt */
2a418930 824 goto haveidle;
4b5f931b 825 }
bd52bedf 826#if defined(INVARIANTS) && defined(__amd64__)
d5b2d319
MD
827 if ((read_rflags() & PSL_I) == 0) {
828 cpu_enable_intr();
829 panic("lwkt_switch() called with interrupts disabled");
830 }
831#endif
b12defdc
MD
832
833 /*
834 * Number iterations so far. After a certain point we switch to
835 * a sorted-address/monitor/mwait version of lwkt_getalltokens()
836 */
837 if (spinning < 0x7FFFFFFF)
838 ++spinning;
839
840#ifdef SMP
841 /*
842 * lwkt_getalltokens() failed in sorted token mode, we can use
843 * monitor/mwait in this case.
844 */
845 if (spinning >= lwkt_spin_loops &&
846 (cpu_mi_feature & CPU_MI_MONITOR) &&
847 lwkt_spin_monitor)
848 {
849 cpu_mmw_pause_int(&gd->gd_reqflags,
850 (gd->gd_reqflags | RQF_SPINNING) &
851 ~RQF_IDLECHECK_WK_MASK);
852 }
853#endif
854
855 /*
856 * We already checked that td is still scheduled so this should be
857 * safe.
858 */
859 splz_check();
860
861 /*
862 * This experimental resequencer is used as a fall-back to reduce
863 * hw cache line contention by placing each core's scheduler into a
864 * time-domain-multplexed slot.
865 *
866 * The resequencer is disabled by default. It's functionality has
867 * largely been superceeded by the token algorithm which limits races
868 * to a subset of cores.
869 *
870 * The resequencer algorithm tends to break down when more than
871 * 20 cores are contending. What appears to happen is that new
872 * tokens can be obtained out of address-sorted order by new cores
873 * while existing cores languish in long delays between retries and
874 * wind up being starved-out of the token acquisition.
875 */
876 if (lwkt_spin_reseq && spinning >= lwkt_spin_reseq) {
877 int cseq = atomic_fetchadd_int(&lwkt_cseq_windex, 1);
878 int oseq;
879
880 while ((oseq = lwkt_cseq_rindex) != cseq) {
881 cpu_ccfence();
882#if 1
883 if (cpu_mi_feature & CPU_MI_MONITOR) {
884 cpu_mmw_pause_int(&lwkt_cseq_rindex, oseq);
885 } else {
886#endif
887 cpu_pause();
888 cpu_lfence();
889#if 1
890 }
8b402283 891#endif
0f0466c0 892 }
b12defdc
MD
893 DELAY(1);
894 atomic_add_int(&lwkt_cseq_rindex, 1);
2a418930 895 }
2a418930 896 /* highest level for(;;) loop */
f1d1c3fa 897 }
8a8d5d85 898
2a418930 899havethread:
b12defdc 900 /*
be71787b
MD
901 * Clear gd_idle_repeat when doing a normal switch to a non-idle
902 * thread.
f9235b6d 903 */
9ac1ee6e 904 ntd->td_wmesg = NULL;
b12defdc 905 ++gd->gd_cnt.v_swtch;
be71787b 906 gd->gd_idle_repeat = 0;
2a418930 907
f9235b6d 908havethread_preempted:
f9235b6d
MD
909 /*
910 * If the new target does not need the MP lock and we are holding it,
911 * release the MP lock. If the new target requires the MP lock we have
912 * already acquired it for the target.
8a8d5d85 913 */
2a418930 914 ;
f9235b6d
MD
915haveidle:
916 KASSERT(ntd->td_critcount,
b5d16701
MD
917 ("priority problem in lwkt_switch %d %d",
918 td->td_critcount, ntd->td_critcount));
919
94f6d86e 920 if (td != ntd) {
cc9b6223
MD
921 /*
922 * Execute the actual thread switch operation. This function
923 * returns to the current thread and returns the previous thread
924 * (which may be different from the thread we switched to).
925 *
926 * We are responsible for marking ntd as TDF_RUNNING.
927 */
121f93bc 928 KKASSERT((ntd->td_flags & TDF_RUNNING) == 0);
94f6d86e 929 ++switch_count;
a1f0fb66 930 KTR_LOG(ctxsw_sw, gd->gd_cpuid, ntd);
cc9b6223
MD
931 ntd->td_flags |= TDF_RUNNING;
932 lwkt_switch_return(td->td_switch(ntd));
933 /* ntd invalid, td_switch() can return a different thread_t */
94f6d86e 934 }
b12defdc 935
b12defdc 936 /*
54341a3b 937 * catch-all. XXX is this strictly needed?
b12defdc
MD
938 */
939 splz_check();
54341a3b 940
37af14fe
MD
941 /* NOTE: current cpu may have changed after switch */
942 crit_exit_quick(td);
8ad65e08
MD
943}
944
cc9b6223
MD
945/*
946 * Called by assembly in the td_switch (thread restore path) for thread
947 * bootstrap cases which do not 'return' to lwkt_switch().
948 */
949void
950lwkt_switch_return(thread_t otd)
951{
952#ifdef SMP
953 globaldata_t rgd;
954
955 /*
956 * Check if otd was migrating. Now that we are on ntd we can finish
957 * up the migration. This is a bit messy but it is the only place
958 * where td is known to be fully descheduled.
959 *
960 * We can only activate the migration if otd was migrating but not
961 * held on the cpu due to a preemption chain. We still have to
962 * clear TDF_RUNNING on the old thread either way.
963 *
964 * We are responsible for clearing the previously running thread's
965 * TDF_RUNNING.
966 */
967 if ((rgd = otd->td_migrate_gd) != NULL &&
968 (otd->td_flags & TDF_PREEMPT_LOCK) == 0) {
969 KKASSERT((otd->td_flags & (TDF_MIGRATING | TDF_RUNNING)) ==
970 (TDF_MIGRATING | TDF_RUNNING));
971 otd->td_migrate_gd = NULL;
972 otd->td_flags &= ~TDF_RUNNING;
973 lwkt_send_ipiq(rgd, lwkt_setcpu_remote, otd);
974 } else {
975 otd->td_flags &= ~TDF_RUNNING;
976 }
977#else
978 otd->td_flags &= ~TDF_RUNNING;
979#endif
980}
981
b68b7282 982/*
96728c05 983 * Request that the target thread preempt the current thread. Preemption
54341a3b
MD
984 * can only occur if our only critical section is the one that we were called
985 * with, the relative priority of the target thread is higher, and the target
986 * thread holds no tokens. This also only works if we are not holding any
987 * spinlocks (obviously).
96728c05
MD
988 *
989 * THE CALLER OF LWKT_PREEMPT() MUST BE IN A CRITICAL SECTION. Typically
990 * this is called via lwkt_schedule() through the td_preemptable callback.
f9235b6d 991 * critcount is the managed critical priority that we should ignore in order
96728c05
MD
992 * to determine whether preemption is possible (aka usually just the crit
993 * priority of lwkt_schedule() itself).
b68b7282 994 *
54341a3b
MD
995 * Preemption is typically limited to interrupt threads.
996 *
997 * Operation works in a fairly straight-forward manner. The normal
998 * scheduling code is bypassed and we switch directly to the target
999 * thread. When the target thread attempts to block or switch away
1000 * code at the base of lwkt_switch() will switch directly back to our
1001 * thread. Our thread is able to retain whatever tokens it holds and
1002 * if the target needs one of them the target will switch back to us
1003 * and reschedule itself normally.
b68b7282
MD
1004 */
1005void
f9235b6d 1006lwkt_preempt(thread_t ntd, int critcount)
b68b7282 1007{
46a3f46d 1008 struct globaldata *gd = mycpu;
cc9b6223 1009 thread_t xtd;
0a3f9b47 1010 thread_t td;
2d910aaf 1011 int save_gd_intr_nesting_level;
b68b7282 1012
26a0694b 1013 /*
96728c05
MD
1014 * The caller has put us in a critical section. We can only preempt
1015 * if the caller of the caller was not in a critical section (basically
f9235b6d 1016 * a local interrupt), as determined by the 'critcount' parameter. We
47737962 1017 * also can't preempt if the caller is holding any spinlocks (even if
d666840a 1018 * he isn't in a critical section). This also handles the tokens test.
96728c05
MD
1019 *
1020 * YYY The target thread must be in a critical section (else it must
1021 * inherit our critical section? I dunno yet).
26a0694b 1022 */
f9235b6d 1023 KASSERT(ntd->td_critcount, ("BADCRIT0 %d", ntd->td_pri));
26a0694b 1024
b12defdc 1025 td = gd->gd_curthread;
fbc024e4
MD
1026 if (preempt_enable == 0) {
1027 ++preempt_miss;
1028 return;
1029 }
f9235b6d 1030 if (ntd->td_pri <= td->td_pri) {
57c254db
MD
1031 ++preempt_miss;
1032 return;
1033 }
f9235b6d 1034 if (td->td_critcount > critcount) {
96728c05
MD
1035 ++preempt_miss;
1036 return;
1037 }
1038#ifdef SMP
121f93bc
MD
1039 if (td->td_cscount) {
1040 ++preempt_miss;
1041 return;
1042 }
46a3f46d 1043 if (ntd->td_gd != gd) {
96728c05
MD
1044 ++preempt_miss;
1045 return;
1046 }
1047#endif
41a01a4d 1048 /*
77912481
MD
1049 * We don't have to check spinlocks here as they will also bump
1050 * td_critcount.
d3d1cbc8
MD
1051 *
1052 * Do not try to preempt if the target thread is holding any tokens.
1053 * We could try to acquire the tokens but this case is so rare there
1054 * is no need to support it.
41a01a4d 1055 */
77912481
MD
1056 KKASSERT(gd->gd_spinlocks_wr == 0);
1057
3b998fa9 1058 if (TD_TOKS_HELD(ntd)) {
d3d1cbc8 1059 ++preempt_miss;
d3d1cbc8
MD
1060 return;
1061 }
26a0694b
MD
1062 if (td == ntd || ((td->td_flags | ntd->td_flags) & TDF_PREEMPT_LOCK)) {
1063 ++preempt_weird;
1064 return;
1065 }
1066 if (ntd->td_preempted) {
4b5f931b 1067 ++preempt_hit;
26a0694b 1068 return;
b68b7282 1069 }
da0b0e8b 1070 KKASSERT(gd->gd_processing_ipiq == 0);
26a0694b 1071
8ec60c3f
MD
1072 /*
1073 * Since we are able to preempt the current thread, there is no need to
1074 * call need_lwkt_resched().
2d910aaf
MD
1075 *
1076 * We must temporarily clear gd_intr_nesting_level around the switch
1077 * since switchouts from the target thread are allowed (they will just
1078 * return to our thread), and since the target thread has its own stack.
cc9b6223
MD
1079 *
1080 * A preemption must switch back to the original thread, assert the
1081 * case.
8ec60c3f 1082 */
26a0694b
MD
1083 ++preempt_hit;
1084 ntd->td_preempted = td;
1085 td->td_flags |= TDF_PREEMPT_LOCK;
a1f0fb66 1086 KTR_LOG(ctxsw_pre, gd->gd_cpuid, ntd);
2d910aaf
MD
1087 save_gd_intr_nesting_level = gd->gd_intr_nesting_level;
1088 gd->gd_intr_nesting_level = 0;
121f93bc
MD
1089
1090 KKASSERT((ntd->td_flags & TDF_RUNNING) == 0);
cc9b6223
MD
1091 ntd->td_flags |= TDF_RUNNING;
1092 xtd = td->td_switch(ntd);
1093 KKASSERT(xtd == ntd);
1094 lwkt_switch_return(xtd);
2d910aaf 1095 gd->gd_intr_nesting_level = save_gd_intr_nesting_level;
b9eb1c19 1096
26a0694b
MD
1097 KKASSERT(ntd->td_preempted && (td->td_flags & TDF_PREEMPT_DONE));
1098 ntd->td_preempted = NULL;
1099 td->td_flags &= ~(TDF_PREEMPT_LOCK|TDF_PREEMPT_DONE);
b68b7282
MD
1100}
1101
f1d1c3fa 1102/*
faaeffac 1103 * Conditionally call splz() if gd_reqflags indicates work is pending.
4a28fe22
MD
1104 * This will work inside a critical section but not inside a hard code
1105 * section.
ef0fdad1 1106 *
f1d1c3fa
MD
1107 * (self contained on a per cpu basis)
1108 */
1109void
faaeffac 1110splz_check(void)
f1d1c3fa 1111{
7966cb69
MD
1112 globaldata_t gd = mycpu;
1113 thread_t td = gd->gd_curthread;
ef0fdad1 1114
4a28fe22
MD
1115 if ((gd->gd_reqflags & RQF_IDLECHECK_MASK) &&
1116 gd->gd_intr_nesting_level == 0 &&
1117 td->td_nest_count < 2)
1118 {
f1d1c3fa 1119 splz();
4a28fe22
MD
1120 }
1121}
1122
1123/*
1124 * This version is integrated into crit_exit, reqflags has already
1125 * been tested but td_critcount has not.
1126 *
1127 * We only want to execute the splz() on the 1->0 transition of
1128 * critcount and not in a hard code section or if too deeply nested.
925040f2
MD
1129 *
1130 * NOTE: gd->gd_spinlocks_wr is implied to be 0 when td_critcount is 0.
4a28fe22
MD
1131 */
1132void
1133lwkt_maybe_splz(thread_t td)
1134{
1135 globaldata_t gd = td->td_gd;
1136
1137 if (td->td_critcount == 0 &&
1138 gd->gd_intr_nesting_level == 0 &&
1139 td->td_nest_count < 2)
1140 {
1141 splz();
1142 }
f1d1c3fa
MD
1143}
1144
e6546af9
MD
1145/*
1146 * Drivers which set up processing co-threads can call this function to
1147 * run the co-thread at a higher priority and to allow it to preempt
1148 * normal threads.
1149 */
1150void
1151lwkt_set_interrupt_support_thread(void)
1152{
1153 thread_t td = curthread;
1154
1155 lwkt_setpri_self(TDPRI_INT_SUPPORT);
1156 td->td_flags |= TDF_INTTHREAD;
1157 td->td_preemptable = lwkt_preempt;
1158}
1159
1160
8ad65e08 1161/*
f9235b6d
MD
1162 * This function is used to negotiate a passive release of the current
1163 * process/lwp designation with the user scheduler, allowing the user
1164 * scheduler to schedule another user thread. The related kernel thread
1165 * (curthread) continues running in the released state.
8ad65e08
MD
1166 */
1167void
f9235b6d 1168lwkt_passive_release(struct thread *td)
8ad65e08 1169{
f9235b6d
MD
1170 struct lwp *lp = td->td_lwp;
1171
1172 td->td_release = NULL;
1173 lwkt_setpri_self(TDPRI_KERN_USER);
1174 lp->lwp_proc->p_usched->release_curproc(lp);
f1d1c3fa
MD
1175}
1176
f9235b6d 1177
3824f392 1178/*
d2d8515b
MD
1179 * This implements a LWKT yield, allowing a kernel thread to yield to other
1180 * kernel threads at the same or higher priority. This function can be
1181 * called in a tight loop and will typically only yield once per tick.
f9235b6d 1182 *
d2d8515b
MD
1183 * Most kernel threads run at the same priority in order to allow equal
1184 * sharing.
f9235b6d
MD
1185 *
1186 * (self contained on a per cpu basis)
3824f392
MD
1187 */
1188void
f9235b6d 1189lwkt_yield(void)
3824f392 1190{
f9235b6d
MD
1191 globaldata_t gd = mycpu;
1192 thread_t td = gd->gd_curthread;
3824f392 1193
f9235b6d
MD
1194 if ((gd->gd_reqflags & RQF_IDLECHECK_MASK) && td->td_nest_count < 2)
1195 splz();
85946b6c 1196 if (lwkt_resched_wanted()) {
f9235b6d
MD
1197 lwkt_schedule_self(curthread);
1198 lwkt_switch();
f9235b6d 1199 }
3824f392
MD
1200}
1201
40504122
MD
1202/*
1203 * The quick version processes pending interrupts and higher-priority
1204 * LWKT threads but will not round-robin same-priority LWKT threads.
1205 */
1206void
1207lwkt_yield_quick(void)
1208{
1209 globaldata_t gd = mycpu;
1210 thread_t td = gd->gd_curthread;
1211
1212 if ((gd->gd_reqflags & RQF_IDLECHECK_MASK) && td->td_nest_count < 2)
1213 splz();
1214 if (lwkt_resched_wanted()) {
1215 if (TAILQ_FIRST(&gd->gd_tdrunq) == td) {
1216 clear_lwkt_resched();
1217 } else {
1218 lwkt_schedule_self(curthread);
1219 lwkt_switch();
1220 }
1221 }
1222}
1223
3824f392 1224/*
f9235b6d
MD
1225 * This yield is designed for kernel threads with a user context.
1226 *
1227 * The kernel acting on behalf of the user is potentially cpu-bound,
1228 * this function will efficiently allow other threads to run and also
1229 * switch to other processes by releasing.
3824f392
MD
1230 *
1231 * The lwkt_user_yield() function is designed to have very low overhead
1232 * if no yield is determined to be needed.
1233 */
1234void
1235lwkt_user_yield(void)
1236{
f9235b6d
MD
1237 globaldata_t gd = mycpu;
1238 thread_t td = gd->gd_curthread;
1239
1240 /*
1241 * Always run any pending interrupts in case we are in a critical
1242 * section.
1243 */
1244 if ((gd->gd_reqflags & RQF_IDLECHECK_MASK) && td->td_nest_count < 2)
1245 splz();
3824f392 1246
3824f392 1247 /*
f9235b6d
MD
1248 * Switch (which forces a release) if another kernel thread needs
1249 * the cpu, if userland wants us to resched, or if our kernel
1250 * quantum has run out.
3824f392 1251 */
f9235b6d 1252 if (lwkt_resched_wanted() ||
85946b6c 1253 user_resched_wanted())
f9235b6d 1254 {
3824f392 1255 lwkt_switch();
3824f392
MD
1256 }
1257
f9235b6d 1258#if 0
3824f392 1259 /*
f9235b6d
MD
1260 * Reacquire the current process if we are released.
1261 *
1262 * XXX not implemented atm. The kernel may be holding locks and such,
1263 * so we want the thread to continue to receive cpu.
3824f392 1264 */
f9235b6d
MD
1265 if (td->td_release == NULL && lp) {
1266 lp->lwp_proc->p_usched->acquire_curproc(lp);
1267 td->td_release = lwkt_passive_release;
1268 lwkt_setpri_self(TDPRI_USER_NORM);
3824f392 1269 }
f9235b6d 1270#endif
b9eb1c19
MD
1271}
1272
8ad65e08 1273/*
f1d1c3fa
MD
1274 * Generic schedule. Possibly schedule threads belonging to other cpus and
1275 * deal with threads that might be blocked on a wait queue.
1276 *
0a3f9b47
MD
1277 * We have a little helper inline function which does additional work after
1278 * the thread has been enqueued, including dealing with preemption and
1279 * setting need_lwkt_resched() (which prevents the kernel from returning
1280 * to userland until it has processed higher priority threads).
6330a558
MD
1281 *
1282 * It is possible for this routine to be called after a failed _enqueue
1283 * (due to the target thread migrating, sleeping, or otherwise blocked).
1284 * We have to check that the thread is actually on the run queue!
8ad65e08 1285 */
0a3f9b47
MD
1286static __inline
1287void
85946b6c 1288_lwkt_schedule_post(globaldata_t gd, thread_t ntd, int ccount)
0a3f9b47 1289{
6330a558 1290 if (ntd->td_flags & TDF_RUNQ) {
85946b6c 1291 if (ntd->td_preemptable) {
f9235b6d 1292 ntd->td_preemptable(ntd, ccount); /* YYY +token */
6330a558 1293 }
0a3f9b47
MD
1294 }
1295}
1296
361d01dd 1297static __inline
8ad65e08 1298void
85946b6c 1299_lwkt_schedule(thread_t td)
8ad65e08 1300{
37af14fe
MD
1301 globaldata_t mygd = mycpu;
1302
cf709dd2
MD
1303 KASSERT(td != &td->td_gd->gd_idlethread,
1304 ("lwkt_schedule(): scheduling gd_idlethread is illegal!"));
cfaeae2a 1305 KKASSERT((td->td_flags & TDF_MIGRATING) == 0);
37af14fe 1306 crit_enter_gd(mygd);
4643740a
MD
1307 KKASSERT(td->td_lwp == NULL ||
1308 (td->td_lwp->lwp_mpflags & LWP_MP_ONRUNQ) == 0);
1309
37af14fe 1310 if (td == mygd->gd_curthread) {
f1d1c3fa
MD
1311 _lwkt_enqueue(td);
1312 } else {
f1d1c3fa 1313 /*
7cd8d145
MD
1314 * If we own the thread, there is no race (since we are in a
1315 * critical section). If we do not own the thread there might
1316 * be a race but the target cpu will deal with it.
f1d1c3fa 1317 */
0f7a3396 1318#ifdef SMP
7cd8d145 1319 if (td->td_gd == mygd) {
9d265729 1320 _lwkt_enqueue(td);
85946b6c 1321 _lwkt_schedule_post(mygd, td, 1);
f1d1c3fa 1322 } else {
e381e77c 1323 lwkt_send_ipiq3(td->td_gd, lwkt_schedule_remote, td, 0);
7cd8d145 1324 }
0f7a3396 1325#else
7cd8d145 1326 _lwkt_enqueue(td);
85946b6c 1327 _lwkt_schedule_post(mygd, td, 1);
0f7a3396 1328#endif
8ad65e08 1329 }
37af14fe 1330 crit_exit_gd(mygd);
8ad65e08
MD
1331}
1332
361d01dd
MD
1333void
1334lwkt_schedule(thread_t td)
1335{
85946b6c 1336 _lwkt_schedule(td);
361d01dd
MD
1337}
1338
1339void
85946b6c 1340lwkt_schedule_noresched(thread_t td) /* XXX not impl */
361d01dd 1341{
85946b6c 1342 _lwkt_schedule(td);
361d01dd
MD
1343}
1344
88ebb169
SW
1345#ifdef SMP
1346
e381e77c
MD
1347/*
1348 * When scheduled remotely if frame != NULL the IPIQ is being
1349 * run via doreti or an interrupt then preemption can be allowed.
1350 *
1351 * To allow preemption we have to drop the critical section so only
1352 * one is present in _lwkt_schedule_post.
1353 */
1354static void
1355lwkt_schedule_remote(void *arg, int arg2, struct intrframe *frame)
1356{
1357 thread_t td = curthread;
1358 thread_t ntd = arg;
1359
1360 if (frame && ntd->td_preemptable) {
1361 crit_exit_noyield(td);
85946b6c 1362 _lwkt_schedule(ntd);
e381e77c
MD
1363 crit_enter_quick(td);
1364 } else {
85946b6c 1365 _lwkt_schedule(ntd);
e381e77c
MD
1366 }
1367}
1368
d9eea1a5 1369/*
52eedfb5
MD
1370 * Thread migration using a 'Pull' method. The thread may or may not be
1371 * the current thread. It MUST be descheduled and in a stable state.
1372 * lwkt_giveaway() must be called on the cpu owning the thread.
1373 *
1374 * At any point after lwkt_giveaway() is called, the target cpu may
1375 * 'pull' the thread by calling lwkt_acquire().
1376 *
ae8e83e6
MD
1377 * We have to make sure the thread is not sitting on a per-cpu tsleep
1378 * queue or it will blow up when it moves to another cpu.
1379 *
52eedfb5 1380 * MPSAFE - must be called under very specific conditions.
d9eea1a5 1381 */
52eedfb5
MD
1382void
1383lwkt_giveaway(thread_t td)
1384{
3b4192fb 1385 globaldata_t gd = mycpu;
52eedfb5 1386
3b4192fb
MD
1387 crit_enter_gd(gd);
1388 if (td->td_flags & TDF_TSLEEPQ)
1389 tsleep_remove(td);
1390 KKASSERT(td->td_gd == gd);
1391 TAILQ_REMOVE(&gd->gd_tdallq, td, td_allq);
1392 td->td_flags |= TDF_MIGRATING;
1393 crit_exit_gd(gd);
52eedfb5
MD
1394}
1395
a2a5ad0d
MD
1396void
1397lwkt_acquire(thread_t td)
1398{
37af14fe
MD
1399 globaldata_t gd;
1400 globaldata_t mygd;
cc9b6223 1401 int retry = 10000000;
a2a5ad0d 1402
52eedfb5 1403 KKASSERT(td->td_flags & TDF_MIGRATING);
a2a5ad0d 1404 gd = td->td_gd;
37af14fe 1405 mygd = mycpu;
52eedfb5 1406 if (gd != mycpu) {
35238fa5 1407 cpu_lfence();
52eedfb5 1408 KKASSERT((td->td_flags & TDF_RUNQ) == 0);
37af14fe 1409 crit_enter_gd(mygd);
cfaeae2a 1410 DEBUG_PUSH_INFO("lwkt_acquire");
df910c23
MD
1411 while (td->td_flags & (TDF_RUNNING|TDF_PREEMPT_LOCK)) {
1412#ifdef SMP
1413 lwkt_process_ipiq();
1414#endif
52eedfb5 1415 cpu_lfence();
cc9b6223
MD
1416 if (--retry == 0) {
1417 kprintf("lwkt_acquire: stuck: td %p td->td_flags %08x\n",
1418 td, td->td_flags);
1419 retry = 10000000;
1420 }
df910c23 1421 }
cfaeae2a 1422 DEBUG_POP_INFO();
562273ea 1423 cpu_mfence();
37af14fe 1424 td->td_gd = mygd;
52eedfb5
MD
1425 TAILQ_INSERT_TAIL(&mygd->gd_tdallq, td, td_allq);
1426 td->td_flags &= ~TDF_MIGRATING;
1427 crit_exit_gd(mygd);
1428 } else {
1429 crit_enter_gd(mygd);
1430 TAILQ_INSERT_TAIL(&mygd->gd_tdallq, td, td_allq);
1431 td->td_flags &= ~TDF_MIGRATING;
37af14fe 1432 crit_exit_gd(mygd);
a2a5ad0d
MD
1433 }
1434}
1435
52eedfb5
MD
1436#endif
1437
f1d1c3fa
MD
1438/*
1439 * Generic deschedule. Descheduling threads other then your own should be
1440 * done only in carefully controlled circumstances. Descheduling is
1441 * asynchronous.
1442 *
1443 * This function may block if the cpu has run out of messages.
8ad65e08
MD
1444 */
1445void
1446lwkt_deschedule(thread_t td)
1447{
f1d1c3fa 1448 crit_enter();
b8a98473 1449#ifdef SMP
f1d1c3fa
MD
1450 if (td == curthread) {
1451 _lwkt_dequeue(td);
1452 } else {
a72187e9 1453 if (td->td_gd == mycpu) {
f1d1c3fa
MD
1454 _lwkt_dequeue(td);
1455 } else {
b8a98473 1456 lwkt_send_ipiq(td->td_gd, (ipifunc1_t)lwkt_deschedule, td);
f1d1c3fa
MD
1457 }
1458 }
b8a98473
MD
1459#else
1460 _lwkt_dequeue(td);
1461#endif
f1d1c3fa
MD
1462 crit_exit();
1463}
1464
4b5f931b
MD
1465/*
1466 * Set the target thread's priority. This routine does not automatically
1467 * switch to a higher priority thread, LWKT threads are not designed for
1468 * continuous priority changes. Yield if you want to switch.
4b5f931b
MD
1469 */
1470void
1471lwkt_setpri(thread_t td, int pri)
1472{
f9235b6d
MD
1473 if (td->td_pri != pri) {
1474 KKASSERT(pri >= 0);
1475 crit_enter();
1476 if (td->td_flags & TDF_RUNQ) {
d2d8515b 1477 KKASSERT(td->td_gd == mycpu);
f9235b6d
MD
1478 _lwkt_dequeue(td);
1479 td->td_pri = pri;
1480 _lwkt_enqueue(td);
1481 } else {
1482 td->td_pri = pri;
1483 }
1484 crit_exit();
26a0694b 1485 }
26a0694b
MD
1486}
1487
03bd0a5e
MD
1488/*
1489 * Set the initial priority for a thread prior to it being scheduled for
1490 * the first time. The thread MUST NOT be scheduled before or during
1491 * this call. The thread may be assigned to a cpu other then the current
1492 * cpu.
1493 *
1494 * Typically used after a thread has been created with TDF_STOPPREQ,
1495 * and before the thread is initially scheduled.
1496 */
1497void
1498lwkt_setpri_initial(thread_t td, int pri)
1499{
1500 KKASSERT(pri >= 0);
1501 KKASSERT((td->td_flags & TDF_RUNQ) == 0);
f9235b6d 1502 td->td_pri = pri;
03bd0a5e
MD
1503}
1504
26a0694b
MD
1505void
1506lwkt_setpri_self(int pri)
1507{
1508 thread_t td = curthread;
1509
4b5f931b
MD
1510 KKASSERT(pri >= 0 && pri <= TDPRI_MAX);
1511 crit_enter();
1512 if (td->td_flags & TDF_RUNQ) {
1513 _lwkt_dequeue(td);
f9235b6d 1514 td->td_pri = pri;
4b5f931b
MD
1515 _lwkt_enqueue(td);
1516 } else {
f9235b6d 1517 td->td_pri = pri;
4b5f931b
MD
1518 }
1519 crit_exit();
1520}
1521
f9235b6d 1522/*
85946b6c 1523 * hz tick scheduler clock for LWKT threads
f9235b6d
MD
1524 */
1525void
85946b6c 1526lwkt_schedulerclock(thread_t td)
f9235b6d 1527{
85946b6c
MD
1528 globaldata_t gd = td->td_gd;
1529 thread_t xtd;
2a418930 1530
85946b6c
MD
1531 if (TAILQ_FIRST(&gd->gd_tdrunq) == td) {
1532 /*
1533 * If the current thread is at the head of the runq shift it to the
1534 * end of any equal-priority threads and request a LWKT reschedule
1535 * if it moved.
1536 */
1537 xtd = TAILQ_NEXT(td, td_threadq);
1538 if (xtd && xtd->td_pri == td->td_pri) {
1539 TAILQ_REMOVE(&gd->gd_tdrunq, td, td_threadq);
1540 while (xtd && xtd->td_pri == td->td_pri)
1541 xtd = TAILQ_NEXT(xtd, td_threadq);
1542 if (xtd)
1543 TAILQ_INSERT_BEFORE(xtd, td, td_threadq);
1544 else
1545 TAILQ_INSERT_TAIL(&gd->gd_tdrunq, td, td_threadq);
1546 need_lwkt_resched();
f9235b6d 1547 }
85946b6c
MD
1548 } else {
1549 /*
1550 * If we scheduled a thread other than the one at the head of the
1551 * queue always request a reschedule every tick.
1552 */
1553 need_lwkt_resched();
f9235b6d
MD
1554 }
1555}
1556
5d21b981 1557/*
52eedfb5
MD
1558 * Migrate the current thread to the specified cpu.
1559 *
cc9b6223
MD
1560 * This is accomplished by descheduling ourselves from the current cpu
1561 * and setting td_migrate_gd. The lwkt_switch() code will detect that the
1562 * 'old' thread wants to migrate after it has been completely switched out
1563 * and will complete the migration.
1564 *
1565 * TDF_MIGRATING prevents scheduling races while the thread is being migrated.
1566 *
1567 * We must be sure to release our current process designation (if a user
1568 * process) before clearing out any tsleepq we are on because the release
1569 * code may re-add us.
ae8e83e6
MD
1570 *
1571 * We must be sure to remove ourselves from the current cpu's tsleepq
1572 * before potentially moving to another queue. The thread can be on
1573 * a tsleepq due to a left-over tsleep_interlock().
5d21b981 1574 */
5d21b981
MD
1575
1576void
1577lwkt_setcpu_self(globaldata_t rgd)
1578{
1579#ifdef SMP
1580 thread_t td = curthread;
1581
1582 if (td->td_gd != rgd) {
1583 crit_enter_quick(td);
cc9b6223 1584
95858b91
MD
1585 if (td->td_release)
1586 td->td_release(td);
ae8e83e6 1587 if (td->td_flags & TDF_TSLEEPQ)
3b4192fb 1588 tsleep_remove(td);
cc9b6223
MD
1589
1590 /*
1591 * Set TDF_MIGRATING to prevent a spurious reschedule while we are
1592 * trying to deschedule ourselves and switch away, then deschedule
1593 * ourself, remove us from tdallq, and set td_migrate_gd. Finally,
1594 * call lwkt_switch() to complete the operation.
1595 */
5d21b981
MD
1596 td->td_flags |= TDF_MIGRATING;
1597 lwkt_deschedule_self(td);
52eedfb5 1598 TAILQ_REMOVE(&td->td_gd->gd_tdallq, td, td_allq);
cc9b6223 1599 td->td_migrate_gd = rgd;
5d21b981 1600 lwkt_switch();
cc9b6223
MD
1601
1602 /*
1603 * We are now on the target cpu
1604 */
1605 KKASSERT(rgd == mycpu);
52eedfb5 1606 TAILQ_INSERT_TAIL(&rgd->gd_tdallq, td, td_allq);
5d21b981
MD
1607 crit_exit_quick(td);
1608 }
1609#endif
1610}
1611
ecdefdda
MD
1612void
1613lwkt_migratecpu(int cpuid)
1614{
1615#ifdef SMP
1616 globaldata_t rgd;
1617
1618 rgd = globaldata_find(cpuid);
1619 lwkt_setcpu_self(rgd);
1620#endif
1621}
1622
cc9b6223 1623#ifdef SMP
5d21b981
MD
1624/*
1625 * Remote IPI for cpu migration (called while in a critical section so we
cc9b6223
MD
1626 * do not have to enter another one).
1627 *
1628 * The thread (td) has already been completely descheduled from the
1629 * originating cpu and we can simply assert the case. The thread is
1630 * assigned to the new cpu and enqueued.
5d21b981 1631 *
cc9b6223 1632 * The thread will re-add itself to tdallq when it resumes execution.
5d21b981
MD
1633 */
1634static void
1635lwkt_setcpu_remote(void *arg)
1636{
1637 thread_t td = arg;
1638 globaldata_t gd = mycpu;
1639
cc9b6223 1640 KKASSERT((td->td_flags & (TDF_RUNNING|TDF_PREEMPT_LOCK)) == 0);
5d21b981 1641 td->td_gd = gd;
562273ea 1642 cpu_mfence();
5d21b981 1643 td->td_flags &= ~TDF_MIGRATING;
cc9b6223 1644 KKASSERT(td->td_migrate_gd == NULL);
4643740a
MD
1645 KKASSERT(td->td_lwp == NULL ||
1646 (td->td_lwp->lwp_mpflags & LWP_MP_ONRUNQ) == 0);
5d21b981
MD
1647 _lwkt_enqueue(td);
1648}
3d28ff59 1649#endif
5d21b981 1650
553ea3c8 1651struct lwp *
4b5f931b
MD
1652lwkt_preempted_proc(void)
1653{
73e4f7b9 1654 thread_t td = curthread;
4b5f931b
MD
1655 while (td->td_preempted)
1656 td = td->td_preempted;
553ea3c8 1657 return(td->td_lwp);
4b5f931b
MD
1658}
1659
99df837e
MD
1660/*
1661 * Create a kernel process/thread/whatever. It shares it's address space
1662 * with proc0 - ie: kernel only.
1663 *
d2d8515b
MD
1664 * If the cpu is not specified one will be selected. In the future
1665 * specifying a cpu of -1 will enable kernel thread migration between
1666 * cpus.
99df837e
MD
1667 */
1668int
c9e9fb21
MD
1669lwkt_create(void (*func)(void *), void *arg, struct thread **tdp,
1670 thread_t template, int tdflags, int cpu, const char *fmt, ...)
99df837e 1671{
73e4f7b9 1672 thread_t td;
e2565a42 1673 __va_list ap;
99df837e 1674
d3d32139 1675 td = lwkt_alloc_thread(template, LWKT_THREAD_STACK, cpu,
dbcd0c9b 1676 tdflags);
a2a5ad0d
MD
1677 if (tdp)
1678 *tdp = td;
709799ea 1679 cpu_set_thread_handler(td, lwkt_exit, func, arg);
99df837e
MD
1680
1681 /*
1682 * Set up arg0 for 'ps' etc
1683 */
e2565a42 1684 __va_start(ap, fmt);
379210cb 1685 kvsnprintf(td->td_comm, sizeof(td->td_comm), fmt, ap);
e2565a42 1686 __va_end(ap);
99df837e
MD
1687
1688 /*
1689 * Schedule the thread to run
1690 */
4643740a
MD
1691 if (td->td_flags & TDF_NOSTART)
1692 td->td_flags &= ~TDF_NOSTART;
ef0fdad1 1693 else
4643740a 1694 lwkt_schedule(td);
99df837e
MD
1695 return 0;
1696}
1697
1698/*
1699 * Destroy an LWKT thread. Warning! This function is not called when
1700 * a process exits, cpu_proc_exit() directly calls cpu_thread_exit() and
1701 * uses a different reaping mechanism.
1702 */
1703void
1704lwkt_exit(void)
1705{
1706 thread_t td = curthread;
c070746a 1707 thread_t std;
8826f33a 1708 globaldata_t gd;
99df837e 1709
2883d2d8
MD
1710 /*
1711 * Do any cleanup that might block here
1712 */
99df837e 1713 if (td->td_flags & TDF_VERBOSE)
6ea70f76 1714 kprintf("kthread %p %s has exited\n", td, td->td_comm);
f6bf3af1 1715 caps_exit(td);
2883d2d8
MD
1716 biosched_done(td);
1717 dsched_exit_thread(td);
c070746a
MD
1718
1719 /*
1720 * Get us into a critical section to interlock gd_freetd and loop
1721 * until we can get it freed.
1722 *
1723 * We have to cache the current td in gd_freetd because objcache_put()ing
1724 * it would rip it out from under us while our thread is still active.
2af9d75d
MD
1725 *
1726 * We are the current thread so of course our own TDF_RUNNING bit will
1727 * be set, so unlike the lwp reap code we don't wait for it to clear.
c070746a
MD
1728 */
1729 gd = mycpu;
37af14fe 1730 crit_enter_quick(td);
2af9d75d
MD
1731 for (;;) {
1732 if (td->td_refs) {
1733 tsleep(td, 0, "tdreap", 1);
1734 continue;
1735 }
1736 if ((std = gd->gd_freetd) != NULL) {
1737 KKASSERT((std->td_flags & (TDF_RUNNING|TDF_PREEMPT_LOCK)) == 0);
1738 gd->gd_freetd = NULL;
1739 objcache_put(thread_cache, std);
1740 continue;
1741 }
1742 break;
c070746a 1743 }
3b4192fb
MD
1744
1745 /*
1746 * Remove thread resources from kernel lists and deschedule us for
2883d2d8
MD
1747 * the last time. We cannot block after this point or we may end
1748 * up with a stale td on the tsleepq.
eb2adbf5
MD
1749 *
1750 * None of this may block, the critical section is the only thing
1751 * protecting tdallq and the only thing preventing new lwkt_hold()
1752 * thread refs now.
3b4192fb
MD
1753 */
1754 if (td->td_flags & TDF_TSLEEPQ)
1755 tsleep_remove(td);
37af14fe 1756 lwkt_deschedule_self(td);
e56e4dea 1757 lwkt_remove_tdallq(td);
74c9628e 1758 KKASSERT(td->td_refs == 0);
2883d2d8
MD
1759
1760 /*
1761 * Final cleanup
1762 */
1763 KKASSERT(gd->gd_freetd == NULL);
c070746a
MD
1764 if (td->td_flags & TDF_ALLOCATED_THREAD)
1765 gd->gd_freetd = td;
99df837e
MD
1766 cpu_thread_exit();
1767}
1768
e56e4dea
MD
1769void
1770lwkt_remove_tdallq(thread_t td)
1771{
1772 KKASSERT(td->td_gd == mycpu);
1773 TAILQ_REMOVE(&td->td_gd->gd_tdallq, td, td_allq);
1774}
1775
9cf43f91
MD
1776/*
1777 * Code reduction and branch prediction improvements. Call/return
1778 * overhead on modern cpus often degenerates into 0 cycles due to
1779 * the cpu's branch prediction hardware and return pc cache. We
1780 * can take advantage of this by not inlining medium-complexity
1781 * functions and we can also reduce the branch prediction impact
1782 * by collapsing perfectly predictable branches into a single
1783 * procedure instead of duplicating it.
1784 *
1785 * Is any of this noticeable? Probably not, so I'll take the
1786 * smaller code size.
1787 */
1788void
b6468f56 1789crit_exit_wrapper(__DEBUG_CRIT_ARG__)
9cf43f91 1790{
b6468f56 1791 _crit_exit(mycpu __DEBUG_CRIT_PASS_ARG__);
9cf43f91
MD
1792}
1793
2d93b37a
MD
1794void
1795crit_panic(void)
1796{
1797 thread_t td = curthread;
850634cc 1798 int lcrit = td->td_critcount;
2d93b37a 1799
850634cc
AH
1800 td->td_critcount = 0;
1801 panic("td_critcount is/would-go negative! %p %d", td, lcrit);
4a28fe22 1802 /* NOT REACHED */
2d93b37a
MD
1803}
1804
d165e668
MD
1805#ifdef SMP
1806
bd8015ca
MD
1807/*
1808 * Called from debugger/panic on cpus which have been stopped. We must still
1809 * process the IPIQ while stopped, even if we were stopped while in a critical
1810 * section (XXX).
1811 *
1812 * If we are dumping also try to process any pending interrupts. This may
1813 * or may not work depending on the state of the cpu at the point it was
1814 * stopped.
1815 */
1816void
1817lwkt_smp_stopped(void)
1818{
1819 globaldata_t gd = mycpu;
1820
1821 crit_enter_gd(gd);
1822 if (dumping) {
1823 lwkt_process_ipiq();
1824 splz();
1825 } else {
1826 lwkt_process_ipiq();
1827 }
1828 crit_exit_gd(gd);
1829}
1830
d165e668 1831#endif